# How Survey Firms Decide Which Lidar Tasks to Automate

**By Dan Cumberland** · Published August 26, 2026 · Categories: AI Strategy

![Featured image for How Survey Firms Decide Which Lidar Tasks to Automate](https://cms.dancumberlandlabs.com/uploads/featured_image_93b04102b0.webp)

Illustration: Dan Cumberland Labs with Gemini.

> AI lidar point cloud classification accuracy hits 92–99% on 5 tasks — ground, buildings, canopy, power lines, roads. 3 tasks still need manual QA review.

AI lidar classification delivers 92–99% accuracy on five well\-defined tasks\.  Three others remain too hard for full automation\.  Knowing which is which changes how you design your QA workflow— and how much of your team's time you can actually reclaim\.

Lidar point cloud classification assigns each laser return to a category defined by ASPRS LAS 1\.4[1](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-1)— ground, buildings, vegetation, power lines, and so on— across a dataset that can hold hundreds of millions of points per project\.  Most survey firms ask "should I use AI classification?"  The question that actually matters: which of the tasks I run most often can I trust AI to classify, and for which ones will I still be running a manual QA pass?

AI lidar classification doesn't succeed or fail universally\.  It succeeds task by task, and the gap between the tasks it nails and the ones it botches is wide enough to matter in [AI implementation for survey workflows](/services/ai-implementation)\.  All accuracy figures in this article come from benchmark conditions— real\-world performance should be validated on representative local data\.

Start with what AI actually handles well\.

## The 5 Lidar Classification Tasks AI Handles Reliably

AI achieves 90%\+ accuracy on five lidar classification tasks with consistent, documented evidence across multiple independent studies\.  These are the tasks worth automating with confidence\.

Across ground extraction, building detection, tree canopy, power lines, and road surfaces, modern deep learning methods clear 92–99% accuracy under benchmark conditions— and commercial software confirms comparable real\-world performance\.

```html-table
<table><thead><tr><th>Task</th><th>Best Documented Accuracy</th><th>Method</th><th>Source</th></tr></thead><tbody><tr><td>Ground/Bare Earth</td><td>92.2% OA / 94.2% F1</td><td>Ensemble DL</td><td>PMC 11548536 (2024)</td></tr><tr><td>Building Footprints</td><td>94.4% F1</td><td>U-Net3+</td><td>SAI Conference (2022)</td></tr><tr><td>Tree Canopy</td><td>High (multiple studies)</td><td>CNN/Transformer</td><td>Multiple (2024–2025)</td></tr><tr><td>Power Lines</td><td>98%+ precision/recall</td><td>Supervised CNN</td><td>MDPI RS (2017), Springer (2024)</td></tr><tr><td>Road Surfaces</td><td>98.9% depth accuracy</td><td>LiDAR + RGB Fusion (DL)</td><td>ASCE (2024)</td></tr></tbody></table>
```

### Ground/Bare Earth Extraction \(DTM/DEM Generation\)

Ensemble deep learning methods achieve 92\.2% overall accuracy and 94\.2% F1\-score on ground point classification[2](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-2)— and this generalizes across terrain types, including an independent validation on a Netherlands dataset at 95\.3%[3](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-3)\.

What makes ground extraction work is that the problem is geometrically ordered: identify the lowest\-height laser returns across sorted spatial regions, separate them from buildings and vegetation above\.  The binary structure, combined with abundant training data from ISPRS benchmarks and datasets like Semantic3D, gives algorithms a strong foundation\.  Ensemble methods that combine two competing point\-cloud neural network designs— PointNet\+\+ and DGCNN— outperform either model alone, improving accuracy from 80\.9% baseline to 92\.2%\.  The same study documented DEM RMSE reduction from 0\.318–1\.362m down to 0\.273–1\.032m[3](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-3)— meaningful precision improvement for elevation\-sensitive projects\.

Performance holds across terrain types: urban terrain at 93\.6%, mountain at 94\.8%, mixed at 92\.2%[2](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-2)\.  The practical caveat is steep slopes— accuracy decreases measurably on grades above 15°[4](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-4)\.  Dense understory is the other edge case\.  That's a failure mode, covered in the next section\.

### Building Footprint Extraction

Building footprint detection achieves 94\.4% F1\-score \(95\.1% correctness, 93\.7% completeness\) under dense urban conditions using U\-Net3\+ architecture[5](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-5)— a clear automation win for environments where manual extraction would take days\.

Buildings have strong geometric signatures: flat roofs, vertical walls, high\-contrast height separation from surrounding terrain\.  Multiple deep learning architectures work here— Mask R\-CNN, U\-Net3\+, and DeepLabV3\+ all produce competitive results\.  Multiple independent methods arriving at similar accuracy is a stronger signal than one paper's headline number\.  In practical terms, AI\-generated building footprints are production\-ready for typical urban and suburban survey contexts\.

The practical caveats: complex roof shapes may fragment into multiple polygons\.  Touching buildings can confuse edge detection\.  Small structures— outbuildings, covered parking structures— sometimes get missed\.  Light spot\-check QA on complex rooftop areas is still warranted, but the bulk of the work is done\.

### Tree Canopy Classification \(High Vegetation\)

Canopy has no single industry benchmark the way ground or power lines do, but CNN and Transformer architectures consistently beat traditional Random Forest methods on canopy\-level detection across every independent study that's tested it[6](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-6)[7](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-7)[8](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-8)\.  This is the part of forest classification that works\.  The part that doesn't \(understory\) comes in the next section\.

The geometric signature of canopy is strong: dense point returns at height, clear separation from ground level\.  Convolutional neural networks and Transformer\-based architectures both capture this effectively, and CNNs outperform traditional Random Forest methods on woody\-versus\-foliage distinction[8](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-8)\.  For survey firms running vegetation classification as part of corridor mapping, wetland delineation, or site planning work, canopy\-level AI classification is ready\.

One distinction to carry into your workflow: this accuracy applies to canopy, not to the vegetation beneath it\.  Individual small trees near boundaries sometimes get missed, and canopy edge blur is a documented artifact\.  But the canopy layer itself is automation\-ready\.

### Power Line and Utility Wire Detection

Power line classification is among the most consistently accurate tasks in automated lidar: precision and recall both exceed 98%, with multiple independent studies confirming the finding from 2017 through 2024[9](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-9)[10](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-10)\.

The reason is structural\.  Power lines have distinct linear geometry and consistent spatial topology\.  The problem is well\-funded— utility companies have invested heavily in lidar\-based inspection workflows— and supervised CNN approaches substantially outperform unsupervised methods\.  A 2024 study documented conductor and ground wire classification at 98\.05% precision, 99\.00% recall, and 98\.58% F\-score[10](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-10)\.  Commercial tool LiPowerline \(GreenValley International\) reports 95% accuracy on transmission lines and 85% on distribution lines[11](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-11)— flagged as vendor claims, slightly lower than peer\-reviewed benchmarks as expected\.  Esri ArcGIS Pro includes a dedicated GeoAI tool for power line classification in point clouds[12](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-12)\.

The practical caveat: dense urban areas with clutter reduce accuracy\.  Distribution lines are harder than transmission\.  Models built for one region may need retraining for different geographies and terrain types\.

### Road and Pavement Surface Extraction

Road surface extraction reaches 98\.9% depth detection accuracy and 98\.0% area measurement accuracy using deep learning methods paired with LiDAR data[13](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-13)— among the highest task\-specific scores documented in the literature\.

Roads have distinct reflectance properties and flat, continuous geometry with clear spatial boundaries\.  The method fuses LiDAR returns with RGB imagery, and combining those two signal types is what pushes road extraction into the highest tier of task\-specific accuracy documented in the literature[13](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-13)\.  That sensor fusion is also a big part of why road extraction runs production\-ready at scale today\.

Caveats worth tracking: unpaved or severely damaged surfaces may misclassify\.  And distinguishing a road from a large parking surface can require additional contextual reasoning\.  But for standard paved infrastructure— highways, surface streets, bridge approaches— AI road extraction is production\-ready\.

That's the confident list\.  Now for the honest one\.

## The 3 Lidar Classification Tasks Where AI Falls Short

Three lidar classification tasks consistently produce unreliable results, and the reason is structural: each one has a problem current algorithms can't overcome\.  Understanding why matters for how you assign QA effort\.

Dense forest understory, bridge decks, and urban clutter share a common trait: the data itself is the problem\.  Either the sensor can't see it, the class is too rare to train on, or the geometry is too irregular to generalize\.

### Dense Vegetation Understory— The Occlusion Problem

Airborne lidar cannot reliably classify vegetation under dense forest canopy\.  The sensor physically can't collect enough points to work with— a laser geometry problem baked into how airborne pulses interact with canopy\.

Dense canopy blocks airborne laser pulses\.  The understory below receives significantly fewer point returns than the canopy above— often too few for reliable classification\.  What data does reach the understory is sparse and ambiguous— not enough to reliably distinguish low vegetation from ground features\.  As one 2024 study documented: "Understory trees may exhibit lower point density and suffer from missing data caused by severe occlusions\."[6](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-6)  A Springer forestry review from the same year named canopy occlusion the "most critical challenge" in dense forest monitoring[7](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-7)\.

This failure mode affects all classification approaches equally\.  Traditional filters like SMRF and CSF fail for the same reason AI does— the data isn't there to work with\.  ISPRS research documented the vertical bias this creates: ground elevation estimates under dense vegetation show upward bias, from negligible under grass to \+0\.15m under aquatic vegetation[14](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-14)\.  Point cloud completion methods \(such as InceptionFormer, 2026\) are working to recover sparse LiDAR data under canopy— primarily for forestry measurement applications— but classification\-grade solutions for airborne survey workflows remain research\-stage[15](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-15)\.

When it might work:

- Terrestrial or mobile LiDAR \(moves the sensor inside the canopy, solving occlusion\)
- Sparse understory in open\-canopy environments
- Point cloud completion methods \(emerging research, not production\-ready\)

### Bridge Deck Classification— Elevation Confusion

Bridge decks are detectable by shape— AI can find them\.  What AI fails at is classifying their elevation correctly: bridge deck points get labeled as ground, which corrupts the digital elevation model downstream\.

USGS error documentation is explicit: "Points on a bridge deck that have been left classified as ground cause an incomplete removal of the bridge in the DEM\."[16](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-16)  Bridge decks are elevated surfaces at mid\-range height— between ground level and the heights associated with buildings\.  That spatial ambiguity, combined with class imbalance \(bridges are rare in training datasets\), means models consistently underperform on this class\.  A 2025 bridge detection study confirmed: decks were "detected with good performance; however, classification errors and insufficient breaklines remain\."[17](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-17)

The breakline problem compounds the issue\.  Even correctly detected bridges often lack the precise edge breaklines needed for accurate DEM generation\.  Manual breakline editing remains the standard correction workflow\.

When it might work:

- Large bridges with strong, unambiguous geometric signatures
- Models trained specifically on bridge\-rich regional datasets
- Supplemental breakline vectorization tools paired with manual QA

### Complex Urban Clutter— Scaffolding, Construction Debris, and Temporary Structures

AI classifies scaffolding, construction debris, trailers, and temporary urban structures poorly because they're geometrically irregular, rare in training data, and change daily\.

The compounding problem has three parts\.  First: infinite shape variety within a class— "scaffolding" means dozens of structural configurations, and the model has to generalize across all of them\.  Second: low representation in benchmark datasets\.  Semantic3D focuses on stable urban categories like buildings and roads; construction clutter is barely represented\.  Third: temporal volatility\.  A scene trained on Monday looks different by Friday\.

A 2025 safety research paper documented the core challenge: "Certain human\-made objects are difficult to detect because of their variety of shapes, irregularly\-distributed point clouds, and a low number of class samples\."[18](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-18)  Incomplete point coverage compounds this further: "When only a small portion of an object appears in the point cloud, detection models receive incomplete shape information, making classification less stable\."[19](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-19)  Urban detection research documents that close vehicles, power lines, and vegetation account for over 95% of false positive cases in point cloud segmentation[20](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-20)— flagging the cross\-contamination risk in high\-clutter scenes\.

AI still handles the majority of urban clutter points correctly\.  But the error rate in active construction zones is significant enough that manual QA is non\-negotiable\.

Now that you know what to trust and what to check— here's how to build the workflow\.

## The Practical Hybrid Workflow— AI First Pass, Then Targeted QA

The industry standard is AI for the first pass, then human QA focused on the three failure zones\.  This is faster than starting from scratch and more reliable than trusting AI blindly\.

Most survey firms ask which software has the best AI classifier\.  The question that matters more: which task failure modes will still need a human review pass, and what that QA workflow actually looks like— the technology has outpaced the workflow design at most firms\.

Hybrid [automating survey workflows](/blog/ai-automation-guide) follows a clear structure:

1. Run AI classification in your platform of choice
2. **High\-confidence / spot\-check:** ground, buildings, canopy, power lines, roads \(92–99% reliable\)
3. **Flag for review:** bridge areas, dense forest edge and understory zones, active construction and high\-clutter urban areas
4. **Run targeted human QA** on flagged zones only— not a full re\-classification

That last point is where the efficiency math changes\.  QA on a first\-pass AI output is targeted work— focused on the three failure zones— rather than a full re\-classification pass across the dataset\.  Industry documentation confirms it: "Processes that used to take days of manual fine\-tuning can now be done in a fraction of the time with higher consistency\."[21](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-21)

```html-table
<table><thead><tr><th>Task</th><th>AI Trust Level</th><th>QA Effort Required</th></tr></thead><tbody><tr><td>Ground Extraction</td><td>High (92–95%)</td><td>Light spot-check</td></tr><tr><td>Building Footprints</td><td>High (94% F1)</td><td>Check complex rooftops</td></tr><tr><td>Tree Canopy</td><td>High</td><td>Verify boundary areas</td></tr><tr><td>Power Lines</td><td>Very High (98%+)</td><td>Confirm in dense clutter</td></tr><tr><td>Road Surfaces</td><td>Very High (98.9%)</td><td>Check unpaved edges</td></tr><tr><td>Dense Understory</td><td>Low</td><td>Manual classification</td></tr><tr><td>Bridge Decks</td><td>Low-Medium</td><td>Manual breakline editing</td></tr><tr><td>Urban Clutter</td><td>Low-Medium</td><td>Manual cleanup passes</td></tr></tbody></table>
```

One method note: Random Forest classification achieves 83–89% F1\-score on ISPRS benchmarks[22](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-22)— adequate for routine work\.  For complex tasks or higher accuracy requirements, deep learning \(CNN, Transformer\) consistently outperforms it\.

Knowing the trust level and QA burden by task only helps if your software actually classifies each of them— which is where tool selection comes in\.

## Software That Uses AI Lidar Classification— and What It Actually Does

TerraSolid, Esri ArcGIS Pro, LiPowerline, and PDAL all offer AI\-based lidar classification\.  What they cover— and don't— varies significantly by task\.

Every major lidar processing platform now uses some form of AI or machine learning for classification\.  What actually differentiates them is which tasks the AI covers and how much human QA it requires\.  [Deciding which AI classification approach fits your firm](/blog/ai-decision-framework-founders) means evaluating tools against the specific task list above, not just their marketing claims\.

```html-table
<table><thead><tr><th>Tool</th><th>Classification Tasks Covered</th><th>Notes</th></tr></thead><tbody><tr><td>TerraSolid (TerraScan)</td><td>Ground, vegetation, buildings, powerlines, infrastructure</td><td>Industry standard; batch macros; no published accuracy benchmarks<sup><a href="#ref-23" class="footnote-ref">23</a></sup></td></tr><tr><td>Esri ArcGIS Pro</td><td>Point cloud classification (PointCNN); dedicated power line tool</td><td>Pre-trained and custom model support; no universal accuracy claims<sup><a href="#ref-24" class="footnote-ref">24</a></sup></td></tr><tr><td>LiPowerline (GreenValley)</td><td>Power line infrastructure (specialized)</td><td>Vendor claims: 95% transmission, 85% distribution<sup><a href="#ref-11" class="footnote-ref">11</a></sup></td></tr><tr><td>PDAL (open source)</td><td>Ground filtering; flexible pipeline integration</td><td>Less turnkey; stronger for custom workflow automation</td></tr></tbody></table>
```

No tool claims universal AI classification across all ASPRS classes\.  Each has documented gaps in the failure zones— bridges, dense vegetation, construction clutter\.

## Frequently Asked Questions About AI Lidar Classification

Common questions practitioners ask about AI lidar classification— answered directly\.

### What accuracy does AI achieve on lidar ground classification?

Modern ensemble deep learning methods achieve 92\.2% overall accuracy and 94\.2% F1\-score on ground classification[2](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-2), improving from approximately 80\.9% with earlier single\-model approaches\.  This result was validated on an independent Netherlands dataset at 95\.3%[3](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-3), indicating cross\-region generalizability\.

### Can AI classify power lines reliably from lidar data?

Yes— power line classification is one of AI's strongest lidar tasks\.  Multiple studies document 98%\+ precision and recall[9](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-9)[10](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-10)\.  Commercial tools like LiPowerline report 95%\+ on transmission lines[11](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-11)\.  Performance decreases in dense urban clutter and for smaller distribution wires\.

### Why does AI struggle with lidar classification in dense forests?

The problem is the sensor, not the algorithm\.  Dense canopy blocks airborne laser pulses, leaving too few points in the understory to classify reliably[6](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-6)[7](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-7)\.  Traditional algorithms fail here for the same reason\.  Terrestrial LiDAR solves the occlusion problem but covers much smaller areas per deployment\.

### How much manual QA is still needed after AI lidar classification?

Typical workflows have AI handling 90–95% of points with confidence\.  [QA costs are often the hidden expense in AI classification projects](/blog/hidden-costs-ai-projects)— targeted manual review covers bridge areas, dense vegetation edges, and urban clutter zones\.  The total is still dramatically faster than full manual classification from scratch\.

### Does one AI model work for all lidar classification tasks?

No\.  Power line models don't generalize to vegetation, and ground models need different training than building detection[2](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-2)[9](/blog/blog-how-survey-firms-decide-which-lidar-tasks-to-automate#ref-9)\.  Task\-specific training— or pre\-trained models for each class— is required\.  There is no universal lidar classifier\.

## Getting the Taxonomy Right Changes the Workflow

AI lidar classification is production\-ready for five core tasks— ground, buildings, canopy, power lines, roads\.  Three tasks still require targeted human review\.  The real shift is from "should I use AI?" to "which zones need QA, and how much effort does each zone actually require?"

The skill in lidar workflows is increasingly workflow design\.  Good AI implementation for survey and AEC work is 10% choosing the right tool and 90% knowing where to direct human judgment\.  The firms that move fastest are the ones who've built clear QA protocols for the failure zones while automating everything else\.

If you're evaluating AI classification tools for your survey or AEC workflows, a structured implementation review can map the right approach to your specific project types\.  Dan Cumberland Labs works with firms on exactly this kind of task\-specific [AI implementation review for survey and AEC workflows](/services/ai-implementation)\.

## References

1. USGS and LidarVisor, "LiDAR Classification Codes: ASPRS LAS Classes 0\-255" \(2022\)— [https://lidarvisor\.com/lidar\-classification\-codes/](https://lidarvisor.com/lidar-classification-codes/)
2. PMC/NIH, "Airborne LiDAR Point Cloud Classification Using Ensemble Learning for DEM Generation" \(2024\)— [https://pmc\.ncbi\.nlm\.nih\.gov/articles/PMC11548536/](https://pmc.ncbi.nlm.nih.gov/articles/PMC11548536/)
3. PMC/NIH, "Ensemble Method— Netherlands AHN Validation Dataset" \(2024\)— [https://pmc\.ncbi\.nlm\.nih\.gov/articles/PMC11548536/](https://pmc.ncbi.nlm.nih.gov/articles/PMC11548536/)
4. ScienceDirect, "Assessment of LiDAR Ground Filtering Algorithms for Determining Ground Surface of Non\-Natural Terrain Overgrown with Forest and Steppe Vegetation" \(2019\)— [https://www\.sciencedirect\.com/science/article/abs/pii/S0263224119309133](https://www.sciencedirect.com/science/article/abs/pii/S0263224119309133)
5. SAI Conference, "Building Footprint Extraction in Dense Area from LiDAR Data Using Mask R\-CNN" \(2022\)— [https://thesai\.org/Publications/ViewPaper?Volume=13&Issue=6&Code=IJACSA&SerialNo=43](https://thesai.org/Publications/ViewPaper?Volume=13&Issue=6&Code=IJACSA&SerialNo=43)
6. PMC/NIH, "Expanding Forest Research with Terrestrial LiDAR Technology" \(2024\)— [https://pmc\.ncbi\.nlm\.nih\.gov/articles/PMC12501281/](https://pmc.ncbi.nlm.nih.gov/articles/PMC12501281/)
7. Springer, "Artificial Intelligence and Terrestrial Point Clouds for Forest Monitoring" \(2024\)— [https://link\.springer\.com/article/10\.1007/s40725\-024\-00234\-4](https://link.springer.com/article/10.1007/s40725-024-00234-4)
8. Springer AI Review, "Deep Learning for Land Use Classification: A Systematic Review of HS\-LiDAR Imagery" \(2025\)— [https://link\.springer\.com/article/10\.1007/s10462\-025\-11265\-z](https://link.springer.com/article/10.1007/s10462-025-11265-z)
9. MDPI Remote Sensing, "Supervised Classification of Power Lines from Airborne LiDAR Data in Urban Areas" \(2017\)— [https://www\.mdpi\.com/2072\-4292/9/8/771](https://www.mdpi.com/2072-4292/9/8/771)
10. Springer Earth Science Informatics, "Powerline Extraction from Aerial and Mobile LiDAR Data Using Deep Learning" \(2024\)— [https://link\.springer\.com/article/10\.1007/s12145\-024\-01310\-w](https://link.springer.com/article/10.1007/s12145-024-01310-w)
11. GreenValley International, "LiPowerline LiDAR Point Cloud Processing" \(2024\)— [https://www\.greenvalleyintl\.com/LiPowerline](https://www.greenvalleyintl.com/LiPowerline) \[vendor claim\]
12. Esri, "Classify Transmission Power Lines in Point Clouds Using Deep Learning" \(2024\)— [https://doc\.esri\.com/en/arcgis\-pro/latest/help/analysis/3d\-analyst/classify\-transmission\-powerlines\-in\-point\-clouds\-using\-deep\-learning\.html](https://doc.esri.com/en/arcgis-pro/latest/help/analysis/3d-analyst/classify-transmission-powerlines-in-point-clouds-using-deep-learning.html)
13. ASCE Journal of Construction Engineering and Management, "LiDAR\-Based Automatic Pavement Distress Detection and Management Using Deep Learning and BIM" \(2024\)— [https://ascelibrary\.org/doi/abs/10\.1061/JCEMD4\.COENG\-14358](https://ascelibrary.org/doi/abs/10.1061/JCEMD4.COENG-14358)
14. ISPRS, "Errors in LiDAR Ground Elevation and Wetland Vegetation Height Estimates" \(2006\)— [https://www\.isprs\.org/proceedings/xxxvi/8\-w2/HOPKINSON\.pdf](https://www.isprs.org/proceedings/xxxvi/8-w2/HOPKINSON.pdf)
15. ScienceDirect, "InceptionFormer: A Deep Learning Framework for UAV LiDAR Point Cloud Completion in Dense Forests" \(2026\)— [https://www\.sciencedirect\.com/science/article/abs/pii/S0034425726001185](https://www.sciencedirect.com/science/article/abs/pii/S0034425726001185)
16. USGS, "LiDAR Classification Codes and Error Documentation— Point Misclassification" \(2022\)— [https://www\.usgs\.gov/media/images/point\-misclassification\-lidar\-error\-6](https://www.usgs.gov/media/images/point-misclassification-lidar-error-6)
17. Tandfonline, "Semi\-Automated Methods for Detection and Measurement of Bridge Assets from Aerial LiDAR" \(2025\)— [https://www\.tandfonline\.com/doi/full/10\.1080/14488353\.2025\.2533670](https://www.tandfonline.com/doi/full/10.1080/14488353.2025.2533670)
18. arXiv, "Safety Assessment of Scaffolding on Construction Site Using AI" \(2025\)— [https://arxiv\.org/pdf/2509\.21368](https://arxiv.org/pdf/2509.21368)
19. iMerit, "LiDAR Perception Challenges in Urban Environments" \(2024\)— [https://imerit\.ai/resources/blog/why\-lidar\-perception\-challenges\-occur\-in\-urban\-settings/](https://imerit.ai/resources/blog/why-lidar-perception-challenges-occur-in-urban-settings/)
20. arXiv, "Detecting the Anomalies in LiDAR Pointcloud" \(2023\)— [https://arxiv\.org/pdf/2308\.00187](https://arxiv.org/pdf/2308.00187)
21. LP360, "Tips for Quality Assurance and Quality Control for LiDAR Data" \(2024\)— [https://www\.lp360\.com/resources/articles/tips\-for\-quality\-assurance\-and\-quality\-control\-qa\-qc\-for\-lidar\-data/](https://www.lp360.com/resources/articles/tips-for-quality-assurance-and-quality-control-qa-qc-for-lidar-data/)
22. MDPI Remote Sensing, "A Supervoxel\-Based Random Forest Method for Robust and Effective Airborne LiDAR Point Cloud Classification" \(2022\)— [https://doi\.org/10\.3390/rs14061516](https://doi.org/10.3390/rs14061516)
23. Terrasolid, "TerraScan Product Documentation" \(2024\)— [https://terrasolid\.com/products/terrascan/](https://terrasolid.com/products/terrascan/)
24. Esri, "Point Cloud Classification Using PointCNN" \(2024\)— [https://developers\.arcgis\.com/python/latest/guide/point\-cloud\-segmentation\-using\-pointcnn/](https://developers.arcgis.com/python/latest/guide/point-cloud-segmentation-using-pointcnn/)


---

**How we made this article:** We use AI in our research and writing so our small team can share more of what we learn. We verify the sources and take responsibility for every article we publish. [Read how we use AI.](https://dancumberlandlabs.com/how-we-use-ai/)

---

## About the author

**Dan Cumberland** — Founder, Dan Cumberland Labs

Dan Cumberland helps engineering and construction firms see where they stand with AI and decide what to build first. He created Pacemark, the AI maturity model behind that work, from research on more than 300 companies.

- Take the assessment: https://pacemark.ai/signal/assessment/?track=aec&utm_source=dcl-site&utm_medium=link&utm_campaign=pacemark-assessment
- Book a call: https://book.dancumberland.com/ai-strategy

---

Source: https://dancumberlandlabs.com/blog/how-survey-firms-decide-which-lidar-tasks-to-automate/